New energy power prediction method and system

Abstract

The invention relates to the technical field of new energy power prediction, and discloses a new energy power prediction method and system. According to the method, through multi-source data preprocessing and correlation screening, key influence factors are focused to construct an initial prediction model; a dynamic mixed meta-heuristic optimization strategy is adopted, global efficient exploration is carried out on the upper layer model, local fine optimization is carried out on the lower layer model, collaborative optimization is carried out on hyper-parameters of the initial prediction model in stages, and finally a target prediction model is formed and used for real-time power prediction. According to the method, through the synergistic effect of the upper-layer model and the lower-layer model of dynamic mixed meta-heuristic optimization, the advantages of different algorithms are combined, and it is ensured that the hyper-parameter optimization process is more likely to converge to a globally optimal solution. The technical problems that an existing single optimization algorithm often has limitation and is prone to falling into local optimum or low in convergence speed are solved.

Description

Technical Field

[0001] The present invention relates to the technical field of new energy power prediction, and in particular to a new energy power prediction method and system. Background Art

[0002] Currently, renewable energy power forecasting methods are primarily categorized into physical, statistical, and artificial intelligence methods. Physical methods are based on numerical weather forecasts and physical models, but they are computationally complex and have limited prediction accuracy. Statistical methods such as ARIMA (Autoregressive Integrated Moving Average) and Kalman filtering, while computationally simple, struggle to capture nonlinear relationships. Artificial intelligence methods, particularly deep learning, excel at handling complex nonlinear relationships and have become a research hotspot. Among deep learning methods, the Long Short-Term Memory (LSTM) network is widely used in time series forecasting due to its ability to learn long-term temporal dependencies. Convolutional Neural Networks (CNNs) excel at extracting spatial features. The CNN-LSTM hybrid network combines the advantages of both, capable of processing both spatial and temporal information. However, the performance of the CNN-LSTM network depends heavily on the selection of hyperparameters. Traditional grid search or random search methods are inefficient and difficult to find the optimal parameter combination.

[0003] However, the performance of traditional deep learning models depends heavily on the settings of hyperparameters. Traditional grid search and random search methods are inefficient and struggle to find the optimal parameter combination. Metaheuristic optimization algorithms such as Particle Swarm Optimization (PSO) and Grey Wolf Optimizer (GWO) can find global optimal solutions in complex search spaces, providing new insights into neural network hyperparameter optimization. However, these single optimization algorithms often have limitations, prone to falling into local optima or slow convergence. Summary of the Invention

[0004] The present invention provides a new energy power prediction method and system, which solves the technical problems that existing single optimization algorithms often have limitations and are prone to falling into local optimality or slow convergence speed.

[0005] A first aspect of the present invention provides a new energy power prediction method, comprising:

[0006] Obtain historical power data of new energy power plants and associated meteorological data and pre-process them to generate multi-source data;

[0007] calculating absolute values ​​of correlations between the multi-source data and the new energy power, and determining a plurality of target multi-source data based on the respective absolute values;

[0008] Inputting each of the target multi-source data into a preset initial prediction model, performing new energy power prediction on each of the target multi-source data using the initial prediction model, and generating an initial prediction result;

[0009] Based on the initial prediction results, the upper layer model and the lower layer model optimized by dynamic hybrid metaheuristic optimization are used to optimize the hyperparameters of the initial prediction model in turn to generate a target prediction model;

[0010] The target prediction model is used to predict the new energy power station's new energy power to generate a target prediction result.

[0011] Optionally, the acquiring of historical power data of the new energy power station and associated meteorological data and preprocessing thereof to generate multi-source data includes:

[0012] Obtaining historical power data and associated meteorological data of a new energy power station; wherein the new energy power station includes wind power generation and photovoltaic power generation, and the meteorological data includes multi-high-speed wind speed, wind direction, and temperature associated with the historical power data of wind power generation, and solar radiation intensity and ambient temperature associated with the historical power data of photovoltaic power generation;

[0013] The historical power data and the associated meteorological data are sequentially subjected to outlier processing, missing value processing and standardization processing to generate multi-source data.

[0014] Optionally, it also includes:

[0015] Determining whether the wind speed of each meteorological data to be measured in the multi-source data is greater than a preset wind speed threshold;

[0016] If yes, determining that the meteorological data to be measured is extreme weather, and marking the meteorological data to be measured as a first extreme weather sample;

[0017] If not, determining whether the photovoltaic solar radiation of each meteorological data to be measured in the multi-source data is less than a preset solar radiation threshold;

[0018] If so, it is determined that the meteorological data to be measured is extreme weather, and the meteorological data to be measured is marked as a second extreme weather sample.

[0019] Optionally, calculating the absolute value of the correlation between the multi-source data and the new energy power, and determining a plurality of target multi-source data based on each of the absolute values, includes:

[0020] Calculating the correlation between the multi-source data and the new energy power based on the Pearson correlation coefficient algorithm;

[0021] determining an absolute value of the correlation based on the correlation;

[0022] Determining whether each of the absolute values ​​is greater than a preset correlation threshold;

[0023] If so, the multi-source data corresponding to the absolute value is determined as the target multi-source data.

[0024] Optionally, the initial prediction model includes a convolutional layer, a long short-term memory layer, a temporal attention layer, and a fully connected layer; inputting each of the target multi-source data into a preset initial prediction model, performing new energy power prediction on each of the target multi-source data using the initial prediction model, and generating an initial prediction result, including:

[0025] Inputting each of the target multi-source data into the convolutional layer of a preset initial prediction model;

[0026] Convolving the features of the target multi-source data through the convolution layer to generate spatial features and input them into the long short-term memory layer and the temporal attention layer;

[0027] Extracting effective components of the time dimension and feature dimension of the spatial feature respectively through the long short-term memory layer, generating linkage features and inputting them into the fully connected layer;

[0028] Extracting the temporal features of the spatial features through the temporal attention layer and inputting them into the fully connected layer;

[0029] The linkage feature and the time series feature are connected through the fully connected layer to generate an initial prediction result.

[0030] Optionally, the initial prediction model further includes an extreme scenario sub-model; and the method further includes:

[0031] When extreme weather exists in the target multi-source data, screening a first extreme weather sample or a second extreme weather sample corresponding to the extreme weather;

[0032] Extracting extreme weather features of the first extreme weather sample or the second extreme weather sample through the extreme scene sub-model and inputting the features into the fully connected layer;

[0033] The linkage feature, the time series feature, and the extreme weather feature are connected through the fully connected layer to generate an initial prediction result.

[0034] Optionally, based on the initial prediction result, optimizing the hyperparameters of the initial prediction model in sequence using an upper model and a lower model optimized by a dynamic hybrid metaheuristic method to generate a target prediction model includes:

[0035] Based on the initial prediction results, the hyperparameter search space of the initial prediction model is input into the upper model of the dynamic hybrid metaheuristic optimization;

[0036] Performing a global exploration of the hyperparameter search space using the starfish optimization algorithm of the upper model to generate multiple hyperparameter optimal domains and inputting them into the lower model optimized by the dynamic hybrid metaheuristic method;

[0037] Optimizing the hyperparameter optimal range in turn by the particle swarm optimization algorithm and the gray wolf optimization algorithm of the lower model to determine the optimal solution;

[0038] The initial prediction model is adjusted according to the target hyperparameters corresponding to the optimal solution to generate a target prediction model.

[0039] A second aspect of the present invention provides a new energy power prediction system, comprising:

[0040] The acquisition module is used to obtain historical power data of new energy power plants and related meteorological data and perform preprocessing to generate multi-source data;

[0041] a calculation module, configured to calculate absolute values ​​of correlations between the multi-source data and the new energy power, and determine a plurality of target multi-source data based on the respective absolute values;

[0042] A prediction module, configured to input each of the target multi-source data into a preset initial prediction model, perform new energy power prediction on each of the target multi-source data using the initial prediction model, and generate an initial prediction result;

[0043] An optimization module is used to optimize the hyperparameters of the initial prediction model in sequence using an upper model and a lower model optimized by a dynamic hybrid metaheuristic method based on the initial prediction result to generate a target prediction model;

[0044] The real-time prediction module is used to predict the new energy power of the new energy power station through the target prediction model and generate a target prediction result.

[0045] A third aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed, implements the new energy power prediction method as described in any one of the above items.

[0046] A fourth aspect of the present invention provides a computer program product, which includes a computer program stored on a non-transitory computer-readable storage medium, and the computer program includes program instructions, wherein when the program instructions are executed by a computer, the computer executes the new energy power prediction method as described in any one of the above items.

[0047] It can be seen from the above technical solutions that the present invention has the following advantages:

[0048] This invention uses multi-source data preprocessing and correlation screening to focus on key influencing factors and construct an initial prediction model. Its core strategy utilizes a dynamic hybrid metaheuristic optimization strategy, using an upper-level model for efficient global exploration and a lower-level model for refined local optimization. The initial prediction model's hyperparameters are collaboratively optimized in stages, ultimately forming a target prediction model for real-time power forecasting. This achieves accurate renewable energy power forecasting driven by multi-source data. The dynamic hybrid metaheuristic optimization approach, where the upper-level model (global exploration) and the lower-level model (local refinement) work synergistically, combining the strengths of different algorithms (such as global search and local development capabilities), effectively avoiding the vulnerability of a single algorithm to local optima and ensuring that the hyperparameter optimization process converges more easily to a globally optimal solution. The upper-level model quickly locates potential optimal regions in the hyperparameter space, reducing ineffective searches; the lower-level model conducts refined optimization within the optimal region, shortening the convergence path. This overcomes the slow convergence speed of a single algorithm and improves overall optimization efficiency. Through multi-source data screening and hierarchical optimization, the target prediction model more accurately captures the dynamic changes in renewable energy power, reducing prediction errors and providing reliable support for renewable energy power plant scheduling and stable power grid operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0050] Figure 1 A flowchart of a new energy power prediction method provided in Example 1 of the present invention;

[0051] Figure 2 A flowchart of a new energy power prediction method provided in the second embodiment of the present invention;

[0052] Figure 3 This is a structural block diagram of a new energy power prediction system provided in Example 3 of the present invention. DETAILED DESCRIPTION

[0053] The embodiments of the present invention provide a new energy power prediction method and system for solving the technical problems that existing single optimization algorithms often have limitations and are prone to falling into local optimality or having a slow convergence speed.

[0054] In order to make the purpose, features, and advantages of the present invention more obvious and easy to understand, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described below are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0055] See also Figure 1 , Figure 1 This is a flowchart of the steps of a new energy power prediction method provided in Example 1 of the present invention.

[0056] The present invention provides a new energy power prediction method, comprising the following steps:

[0057] Step 101: Obtain historical power data of a new energy power station and associated meteorological data and perform preprocessing to generate multi-source data.

[0058] In an embodiment of the present invention, the new energy power station includes wind power generation and photovoltaic power generation.

[0059] Historical power data refers to the actual output power records of new energy power stations such as wind power generation or photovoltaic power generation over a period of time (such as the past 1 year or 3 years). It is the core data that reflects the changes in the power generation capacity of the power station over time.

[0060] Meteorological data associated with historical power data refers to meteorological observation data that matches historical power data in time and space, and is a key factor in explaining power changes. Wind power generation-related meteorological data includes wind speed at multiple heights (such as 10m, 30m, and hub height), wind direction, ambient temperature, humidity, and air pressure. These data directly affect the efficiency of wind turbines in capturing wind energy.

[0061] Photovoltaic power generation-related meteorological data: including solar radiation intensity (direct radiation, scattered radiation), ambient temperature, component temperature, wind speed, humidity, etc. These data determine the photoelectric conversion efficiency of photovoltaic modules.

[0062] Preprocessing refers to the process of cleaning, correcting and standardizing the original collected historical power data and related meteorological data. The purpose is to eliminate data noise, fill in missing information, unify the data format, and provide high-quality input for subsequent model training.

[0063] Multi-source data refers to a comprehensive dataset that, after preprocessing, integrates historical power data and various related meteorological data. For example, multi-source data for wind power generation might include "wind power every 15 minutes for the past three years, plus the corresponding wind speed at 10m, wind speed at 30m, wind direction, temperature, and air pressure." Multi-source data for photovoltaic power generation might include "photovoltaic power every 30 minutes for the past two years, plus the corresponding solar radiation intensity, ambient temperature, module temperature, and wind speed."

[0064] By collecting historical power generation records (historical power data) from renewable energy power plants (wind and photovoltaic power plants) and environmental data affecting power generation (linked meteorological data), and then undergoing pre-processing steps such as cleaning, correction, and standardization, this multi-dimensional, multi-source data is integrated into high-quality, multi-source data. The key is to ensure the integrity, accuracy, and relevance of this data, providing reliable data support for subsequent screening of key features and building predictive models, while preventing noise, missing data, or inconsistencies in the raw data from impacting model performance.

[0065] Step 102: Calculate the absolute value of the correlation between the multi-source data and the new energy power, and determine a plurality of target multi-source data based on the respective absolute values.

[0066] In this embodiment of the present invention, correlation refers to the Pearson correlation coefficient, which is used to quantify the degree of linear association between various meteorological characteristics (such as wind speed and solar radiation intensity) in multi-source data and renewable energy power. Its value range is [-1, 1], with larger absolute values ​​indicating stronger associations. When the absolute value approaches 1, it indicates a significant linear dependence between the meteorological characteristic and power changes; when the absolute value approaches 0, it indicates a weak linear association.

[0067] Target multi-source data refers to the subset of multi-source data retained after correlation screening that has a strong correlation with renewable energy power (i.e., the absolute value of the correlation is greater than a set correlation threshold, such as 0.3). This data is selected from key features of the original multi-source data. It includes both meteorological data with significant impact on power (such as wind speed at hub height for wind turbines and solar radiation intensity for photovoltaics), as well as records of historical power data with strong correlations with meteorological features. These records serve as the core input for subsequent model training.

[0068] The linear correlation strength between each meteorological feature and the new energy power in the multi-source data is calculated based on the Pearson correlation coefficient, and the features whose absolute value of the correlation exceeds the preset correlation threshold are included in the target multi-source data, and the features with weak correlation (such as small pressure fluctuation data with extremely low correlation with power) are eliminated.

[0069] Step 103: Input each target multi-source data into a preset initial prediction model, perform new energy power prediction on each target multi-source data through the initial prediction model, and generate an initial prediction result.

[0070] In an embodiment of the present invention, the initial prediction model refers to a pre-built basic model used to preliminarily learn the correlation between target multi-source data and new energy power. The core architecture includes a long short-term memory network (LSTM) and a temporal attention mechanism.

[0071] Renewable energy power forecasting refers to the process of estimating the output power of renewable energy power stations in the future by learning the inherent laws in the data through an initial prediction model based on target multi-source data (such as historical power, wind speed, solar radiation and other key features).

[0072] The initial prediction result refers to the power estimate output by the initial prediction model after predicting the target multi-source data without hyperparameter optimization. Because model parameters (such as learning rate and number of neurons) have not yet been tuned, this result may have large errors and insufficient stability.

[0073] The target multi-source data that has been relevance-screened (such as wind speed at hub height for wind turbines and solar radiation intensity for photovoltaics) is input into the initial prediction model. The model uses LSTM to learn the temporal patterns in the data (such as the periodic correlation between intraday wind speed changes and power), while focusing on key moment features through the attention mechanism (such as the sudden change in wind speed before a gust of wind). Finally, it outputs the initial prediction result, which is a preliminary estimate of the power value corresponding to the input data.

[0074] Step 104: Based on the initial prediction results, the upper model and the lower model optimized by the dynamic hybrid metaheuristic method are used to optimize the hyperparameters of the initial prediction model in turn to generate a target prediction model.

[0075] In this embodiment of the present invention, dynamic hybrid metaheuristics refers to a method that combines the strengths of multiple metaheuristic algorithms, using hierarchical collaboration and dynamic adjustment strategies to efficiently optimize the hyperparameters of the initial prediction model. Its core concept is "dynamic hybridization": based on the search state during the optimization process (e.g., whether it has reached a stagnation), the search logic of different algorithms is flexibly switched or integrated, balancing global exploration and local development capabilities, thus avoiding the limitations of a single algorithm (such as local optimality and slow convergence).

[0076] The upper-level model, which is the algorithmic module responsible for global rough search in dynamic hybrid metaheuristic optimization, is used to quickly locate potential optimal regions in the hyperparameter space (i.e., regions that may contain the global optimal solution). For example, the Starfish Optimization Algorithm (SFOA) or Particle Swarm Optimization (PSO) algorithm uses its strong exploration capabilities to traverse a large range of hyperparameter spaces, screen out candidate regions worth further search, and reduce inefficient computation.

[0077] The lower-level model refers to the algorithmic module responsible for local fine-tuning in dynamic hybrid metaheuristic optimization. Its role is to conduct a detailed search within the optimal domain identified by the upper-level model, optimizing the specific values ​​of hyperparameters. For example, the Grey Wolf Optimizer (GWO) or the Starfish Optimizer (SFOA) algorithm can be used, leveraging their robust development capabilities to iteratively search for optimal solutions within a small domain, ultimately approaching the global optimal solution. Furthermore, the lower-level model dynamically switches algorithms based on search progress (for example, switching algorithms if no improvement has been achieved after multiple generations), enhancing optimization flexibility.

[0078] The target prediction model refers to a prediction model whose hyperparameters achieve a globally optimal configuration after dynamic hybrid metaheuristic optimization. It retains the core LSTM and attention mechanism architecture of the initial prediction model, but through the optimization of hyperparameters (such as learning rate, number of neurons, and attention weight coefficient), it significantly improves its ability to capture the changing patterns of renewable energy power. Its prediction accuracy and stability surpass those of the initial model, making it directly applicable to power prediction in real-world scenarios.

[0079] Based on the initial prediction results (including information such as error patterns and fitting performance), the direction of hyperparameter optimization is determined (for example, if the initial model has large errors in extreme scenarios, the number of LSTM neurons needs to be optimized). Dynamic hybrid metaheuristic optimization is then initiated: the upper-level model first conducts a global exploration, rapidly narrowing the hyperparameter search range and identifying potential optimal regions. The lower-level model then conducts a detailed optimization within the optimal region, dynamically switching algorithms (e.g., PSO to GWO) to avoid local optimality and gradually adjust hyperparameters to optimal values. Through the coordinated optimization of the upper and lower-level models, the hyperparameter configuration of the initial prediction model reaches a globally optimal state, forming a target prediction model that can be directly used for actual predictions.

[0080] Step 105: Use the target prediction model to predict the new energy power of the new energy power station and generate a target prediction result.

[0081] In this embodiment of the present invention, the target prediction result refers to the estimated output power of a new energy power station over a period of time, output by a target prediction model optimized using a dynamic hybrid metaheuristic method, which analyzes and calculates the plant's real-time input data (such as real-time meteorological data and recent power data). This result is the final prediction generated by the model through hyperparameter optimization and comprehensive learning of multi-source data patterns. It features high accuracy and strong stability, and can directly reflect future trends in power station output (e.g., wind power / photovoltaic output power 1 hour and 24 hours later).

[0082] Real-time data from renewable energy power stations (such as current wind speed, solar radiation intensity, and recent power fluctuations) is input into the target prediction model. The model captures the temporal dependencies of real-time data (such as the impact of current wind speed trends on future power) through an optimized LSTM layer. Combined with a temporal attention mechanism, it focuses on key real-time features (such as sudden gusts of wind and cloud cover), ultimately generating a target prediction result—an accurate estimate of renewable energy output power for a specific period in the future.

[0083] See also Figure 2 , Figure 2 This is a flowchart of the steps of a new energy power prediction method provided in Example 2 of the present invention.

[0084] The present invention provides a new energy power prediction method, comprising the following steps:

[0085] Step 201: Obtain historical power data of a new energy power station and associated meteorological data and perform preprocessing to generate multi-source data.

[0086] Furthermore, step 201 includes the following sub-steps:

[0087] S11. Obtain historical power data of new energy power stations and associated meteorological data; wherein, the new energy power stations include wind power generation and photovoltaic power generation, and the meteorological data include multi-high-speed wind speed, wind direction, and temperature associated with the historical power data of wind power generation, and solar radiation intensity and ambient temperature associated with the historical power data of photovoltaic power generation.

[0088] In this embodiment of the present invention, the actual output power records of wind and photovoltaic power stations over a period of time (e.g., wind power output every 15 minutes for wind power stations and photovoltaic power output every 30 minutes for photovoltaic power stations) are obtained to provide the basis for reflecting the temporal changes in power generation capacity of power stations. Associated meteorological data and historical power data are environmental parameters that match temporally and spatially, and differ depending on the type of power station. Meteorological data associated with wind power generation includes wind speed at multiple heights (e.g., wind speed at 10 m, 30 m, and at the turbine hub height; wind speeds at different heights have different effects on wind energy capture by the turbine), wind direction (affecting the turbine's windward efficiency), and temperature (affecting air density and, in turn, wind power output), which directly determine wind power output. Meteorological data associated with photovoltaic power generation includes solar radiation intensity (the energy source for photovoltaic panels and a key factor affecting photovoltaic power generation) and ambient temperature (excessive temperature reduces the conversion efficiency of photovoltaic modules), which are key drivers of photovoltaic power fluctuations.

[0089] S12. Perform outlier processing, missing value processing, and standardization processing on the historical power data and the associated meteorological data in sequence to generate multi-source data.

[0090] In the embodiment of the present invention, outlier processing adopts the principle of three times the standard deviation to identify outliers and replaces them with the mean of the previous and next moments.

[0091] Missing value imputation processing: linear interpolation method is used to fill missing data.

[0092] Standardization processing uses maximum-minimum standardization to map the data to the [0,1] interval. The specific calculation formula is:

[0093] .

[0094] In the formula, x represents the original number of samples; represents the standardized data sample, and Represent the minimum and maximum values ​​of the data set respectively.

[0095] The historical power data and related meteorological data are preprocessed in sequence according to the methods of outlier processing, missing value filling and standardization to obtain multi-source data.

[0096] Furthermore, the method further comprises the following sub-steps:

[0097] S21 , determining whether the wind speed of each meteorological data to be measured in the multi-source data is greater than a preset wind speed threshold.

[0098] In the embodiment of the present invention, each meteorological data to be measured refers to meteorological data that needs to be judged whether it is extreme weather.

[0099] The preset wind speed threshold is set to 25m / s.

[0100] Use the Isolation Forest algorithm to mark extreme samples, for example, wind speed > 25m / s or photovoltaic solar radiation < 50W / m 2 .

[0101] Determine whether the wind speed of each meteorological data to be measured in all multi-source data is greater than 25m / s.

[0102] S22: If yes, determine that the meteorological data to be measured is extreme weather, and mark the meteorological data to be measured as the first extreme weather sample.

[0103] In the embodiment of the present invention, the first extreme weather sample refers to extreme weather with a wind speed greater than 25 m / s.

[0104] If the wind speed is greater than 25 m / s, the meteorological data to be measured is determined to be extreme weather, and the meteorological data to be measured is marked as the first extreme weather sample.

[0105] S23: If not, determine whether the photovoltaic solar radiation of each meteorological data to be measured in the multi-source data is less than a preset solar radiation threshold.

[0106] In the embodiment of the present invention, the preset solar radiation threshold is 50W / m 2 .

[0107] Determine whether the photovoltaic solar radiation of each meteorological data to be measured in all multi-source data is less than 50W / m 2 .

[0108] S24: If yes, determine that the meteorological data to be measured is extreme weather, and mark the meteorological data to be measured as the second extreme weather sample.

[0109] In the embodiment of the present invention, the second extreme weather sample refers to a situation where the photovoltaic solar radiation is less than 50W / m 2 extreme weather.

[0110] If photovoltaic solar radiation is less than 50W / m 2 , determine that the meteorological data to be tested is extreme weather, and mark the meteorological data to be tested as the second extreme weather sample.

[0111] Step 202: Calculate the correlation between the multi-source data and the new energy power based on the Pearson correlation coefficient algorithm.

[0112] It's important to note that the Pearson correlation coefficient algorithm is used to quantify the linear correlation between various meteorological characteristics (such as wind speed and solar radiation intensity) and renewable energy power in multi-source data. Its value range is [-1, 1], with larger absolute values ​​indicating stronger correlations. When the absolute value approaches 1, it indicates a significant linear dependence between the meteorological characteristic and power changes; when the absolute value approaches 0, the linear correlation is weak.

[0113] The calculation formula of Pearson correlation coefficient is:

[0114]

[0115] In the formula, X and Y represent multi-source data and new energy power respectively. and They represent the corresponding means, n is the number of samples; Represents the i-th sample value of variable X; Represents the i-th sample value of variable Y.

[0116] In the embodiment of the present invention, the Pearson correlation coefficient is used to calculate the correlation between each multi-source data and the new energy power.

[0117] Step 203: Determine the absolute value of the correlation according to the correlation.

[0118] In the embodiment of the present invention, according to the correlation value , the absolute value of the correlation can be obtained .

[0119] Step 204: Determine whether each absolute value is greater than a preset correlation threshold.

[0120] In the embodiment of the present invention, the preset correlation threshold is 0.3.

[0121] Screen for absolute values ​​greater than 0.3.

[0122] Step 205: If yes, the multi-source data corresponding to the absolute value is determined as the target multi-source data.

[0123] In the embodiment of the present invention, when the absolute value is greater than 0.3, the multi-source data corresponding to the absolute value is included in the target multi-source data.

[0124] Step 206: Input each target multi-source data into a preset initial prediction model, perform new energy power prediction on each target multi-source data through the initial prediction model, and generate an initial prediction result.

[0125] Furthermore, the initial prediction model includes a convolutional layer, a long short-term memory layer, a temporal attention layer, and a fully connected layer; step 206 includes the following sub-steps:

[0126] S31. Input each target multi-source data into the convolution layer of the preset initial prediction model.

[0127] In this embodiment of the present invention, the convolution operation formula of the convolution layer (CNN layer) is:

[0128]

[0129] Where, is the jth output feature, is the input sequence, is the convolution kernel weight, b is the bias term, k is the convolution kernel size, and f is the activation function.

[0130] Input each target multi-source data into the convolution layer of the preset initial prediction model in a certain order.

[0131] S32. Convolve the features of the target multi-source data through the convolution layer to generate spatial features and input them into the long short-term memory layer and the temporal attention layer.

[0132] In the embodiment of the present invention, spatial features refer to features obtained by convolving features of target multi-source data through a convolution layer.

[0133] According to the convolution operation formula of the convolution layer, the features of each target multi-source data input into the convolution layer are convolved to obtain spatial features, and the spatial features are input into the long short-term memory layer and the temporal attention layer respectively.

[0134] S33. The effective components of the time dimension and feature dimension of the spatial features are extracted respectively through the long short-term memory layer, and the linkage features are generated and input into the fully connected layer.

[0135] In this embodiment of the present invention, the long short-term memory layer includes a forget gate, an input gate, a candidate value, a state update, an output gate, and a hidden state. The specific calculation formula is:

[0136] Forget Gate:

[0137] ;

[0138] Input Gate:

[0139] ;

[0140] Candidate values:

[0141] ;

[0142] Cell status update:

[0143] ;

[0144] Output gate:

[0145] ;

[0146] Hidden state:

[0147] ;

[0148] Where, is the sigmoid activation function, W is the weight matrix, b is the bias vector, is the hidden state at time t-1, is the input at time t, represents the cell state at time t, represents the cell state at time t-1, represents the candidate cell state at time t.

[0149] The long short-term memory layer (LSTM) extracts effective information from two dimensions through its gating mechanism (forget gate, input gate, output gate, etc.):

[0150] The effective component of the time dimension captures the dynamic changes of spatial features over time. For example, in a wind power scenario, the spatial distribution of wind speeds at different heights continuously changes at times t-1, t, and t+1. LSTM will filter out time series segments that are critical for power prediction (such as the moment when the spatial distribution of wind speeds changes from dispersed to concentrated).

[0151] Feature dimension effective component: Exploits the correlation between different spatial features. For example, in photovoltaic scenarios, the spatial distribution of solar radiation intensity and the spatial distribution of ambient temperature are not independent. LSTM can extract the synergistic variation pattern between the two (such as the characteristic combination of enhanced radiation accompanied by increased temperature).

[0152] The LSTM layer fuses the effective components of the above-mentioned time dimension and feature dimension to generate linkage features. These features not only contain the temporal dynamic laws of spatial features, but also integrate the interactive influence of multi-dimensional features. They can more comprehensively reflect the complex driving mechanism of changes in renewable energy power (such as the combined impact of "temporal changes in wind speed spatial distribution + synergistic effects of temperature characteristics" on wind power).

[0153] Finally, the linkage features are input into the fully connected layer, and the power prediction value is output after nonlinear mapping, providing high-quality feature support for accurate prediction.

[0154] S34. Extract the temporal features of the spatial features through the temporal attention layer and input them into the fully connected layer.

[0155] In this embodiment of the present invention, the temporal attention layer refers to a network component designed to mine key information in the temporal dimension of sequence data. Based on the concept of the attention mechanism, when processing data containing time series characteristics (such as time-varying renewable energy power plant power and meteorological data), it calculates attention weights for features at different time steps. This layer emphasizes key time series segments that have a significant impact on prediction results, suppresses irrelevant or minor time series information, and enhances the model's ability to capture time series dependencies and important temporal patterns. It is often used in conjunction with recurrent neural networks (such as LSTM) and convolutional neural networks for tasks such as time series prediction and sequence classification.

[0156] Through the attention mechanism of the temporal attention layer, weights are calculated for spatial features at different time steps, focusing on those temporal point features that are more critical to the final prediction (such as new energy power prediction) (such as the temporal distribution characteristics of wind speed corresponding to the moment of wind power mutation), screening and strengthening effective temporal information, suppressing redundant and noisy temporal content, and extracting temporal features containing important time correlations and change patterns.

[0157] The time series features extracted and enhanced by the time series attention layer are passed to the fully connected layer as more refined and targeted key information. The fully connected layer then performs subsequent nonlinear transformations, feature fusion, and final prediction output (such as calculation of new energy power forecasts) based on these time series features. This allows the model to utilize this focused time series information to complete prediction tasks, improving the efficiency of time series data utilization and prediction accuracy.

[0158] S35. Connect the linkage features and the time series features through the fully connected layer to generate the initial prediction results.

[0159] In an embodiment of the present invention, the fully connected layer serves as a key module for feature fusion and prediction mapping. It first performs dimensional adaptation and splicing on the linkage features extracted from the spatial features and containing long- and short-term memory interaction information, as well as the temporal features obtained by focusing on key temporal patterns by the temporal attention layer, to open up the information space of the two types of features, allowing the model to simultaneously utilize multi-feature interactions and key temporal laws; subsequently, the spliced ​​fusion features are linearly transformed and nonlinearly activated by the fully connected layer, mapped from the high-dimensional feature space to the prediction target dimension, and generate an initial prediction result that has not been optimized for hyperparameters. This result not only verifies the feasibility of the process, but also provides a basis for error analysis and other aspects for subsequent model optimization. It is a key connection from feature extraction to prediction application.

[0160] Furthermore, the initial prediction model also includes an extreme scenario sub-model; and the method further includes:

[0161] S41. When extreme weather exists in the target multi-source data, the first extreme weather sample or the second extreme weather sample corresponding to the extreme weather is filtered.

[0162] In an embodiment of the present invention, the structure of the extreme scenario sub-model is a simplified CNN-LSTM model, in which the number of convolution kernels and the number of LSTM neurons are reduced.

[0163] When extreme weather is detected in the target multi-source data, the extreme scenario sub-model and the main model (the main model of the initial prediction model) are used for parallel prediction. Therefore, the first extreme weather sample with wind power speed > 25m / s and photovoltaic solar radiation < 50W / m 2 The second most extreme weather sample.

[0164] S42. Extract extreme weather features of the first extreme weather sample or the second extreme weather sample through the extreme scene sub-model and input them into the fully connected layer.

[0165] In the embodiment of the present invention, the extreme weather feature refers to the feature obtained by extracting the first extreme weather sample or the second extreme weather sample through the extreme scene sub-module.

[0166] When faced with the first extreme weather samples (such as extreme strong winds and abnormal wind directions related to wind power generation) or the second extreme weather samples (such as extreme strong radiation and extreme low temperatures related to photovoltaic power generation), the extreme scenario sub-model will extract extreme weather features that reflect the essence of extreme weather and have a key impact on new energy power from these extreme weather samples based on its specific algorithm logic (adapting to the characteristics of extreme scenario data, which may include strengthening the mechanism for capturing outliers and mutation features), such as airflow turbulence characteristics under extreme wind speeds and abnormal energy input characteristics during extreme radiation. These extreme weather features are then input into the fully connected layer and participate in subsequent fusion and prediction calculations together with other conventional features (linkage features, time series features, etc.). This allows the model to improve the accuracy of new energy power prediction under extreme conditions based on the accurately extracted extreme features when dealing with special but significant scenarios such as extreme weather, and improve the overall prediction system's coverage of complex scenarios.

[0167] S43. Connect the linkage features, time series features, and extreme weather features through the fully connected layer to generate the initial prediction results.

[0168] In an embodiment of the present invention, the linkage features containing long- and short-term memory interaction information, the key temporal features focused by temporal attention, and the extreme weather features extracted by the extreme scenario sub-model are dimensionally adapted and spliced ​​through a fully connected layer to integrate the key information captured by each of the three types of features - the linkage features reflect the dynamic interaction of multiple features, the temporal features highlight the rules of important time nodes, and the extreme weather features focus on abnormal patterns in special scenarios; subsequently, through the linear transformation and nonlinear activation of the fully connected layer, the fused high-dimensional features are mapped to the new energy power prediction dimension to generate an initial prediction result without hyperparameter optimization.

[0169] Step 207: Based on the initial prediction results, the upper layer model and the lower layer model optimized by the dynamic hybrid meta-heuristic method are used to optimize the hyperparameters of the initial prediction model in turn to generate a target prediction model.

[0170] Furthermore, step 207 includes the following sub-steps:

[0171] S51. Based on the initial prediction results, the hyperparameter search space of the initial prediction model is input into the upper model of the dynamic hybrid metaheuristic optimization.

[0172] In this embodiment of the present invention, the hyperparameter search space refers to the multidimensional space consisting of the possible ranges of values ​​for all hyperparameters to be optimized during machine learning model optimization. Before model training, a series of hyperparameters (such as the learning rate, number of hidden layer neurons, and regularization coefficient) must be manually set. These parameters cannot be learned by the model itself and must be optimized within a preset space through an algorithm. The search space is typically defined as a range of parameter values ​​(e.g., learning rate ∈ [0.001, 0.1]) or discrete options (e.g., activation function ∈ {ReLU, Sigmoid} or number of LSTM units ∈ [32, 128]; number of CNN convolution kernels ∈ [16, 64]; dropout rate ∈ [0.1, 0.5]; batch size ∈ [16, 64]). The boundaries of the search space must be dynamically adjusted based on domain knowledge, empirical data, and initial model performance.

[0173] Based on the initial prediction results (the power estimates of the unoptimized model), the model's error patterns (such as accuracy in typical scenarios and stability in extreme scenarios) under the current hyperparameter configuration are analyzed. This serves as a basis for defining a reasonable search range for the initial prediction model's hyperparameters (i.e., the hyperparameter search space, encompassing the value boundaries of key parameters such as the learning rate, number of neurons, and dropout rate). This search space is then fed into the upper-level model optimized using a dynamic hybrid metaheuristic. The upper-level model (responsible for global exploration) leverages its strong global search capabilities to rapidly traverse the defined hyperparameter space and initially locate potential sweet spots (regions that may contain the global optimal solution), laying the foundation for local fine-tuning of the lower-level model.

[0174] S52. Globally explore the hyperparameter search space through the starfish optimization algorithm of the upper model, generate multiple hyperparameter optimal domains and input them into the lower model of dynamic hybrid metaheuristic optimization.

[0175] In the embodiment of the present invention, the hyperparameter optimal domain refers to several local areas in the hyperparameter search space that are obtained after preliminary screening by the upper-level model and may contain the global optimal hyperparameter combination.

[0176] The starfish optimization algorithm refers to a heuristic algorithm that simulates the foraging behavior of starfish and has strong global search and adaptive exploration capabilities.

[0177] In a dynamic hybrid metaheuristic optimization framework, the upper model uses the starfish optimization algorithm to globally traverse the hyperparameter search space. By mimicking the efficient foraging strategies of starfish in the vast ocean, this algorithm rapidly explores the multidimensional range of hyperparameter values ​​and identifies local regions where the initial prediction model's performance (e.g., reduced prediction error, improved adaptability to extreme scenarios, etc.) is significantly improved. These regions, known as hyperparameter sweet spots, integrate key information from linkage, temporal, and extreme weather features, focusing on the range of hyperparameter combinations that are crucial for model prediction. The upper model then feeds these sweet spots into the lower model, providing a precise starting point for more refined local optimization (e.g., gradient descent-based parameter fine-tuning). This avoids the inefficiency of blind global search while increasing the probability of finding the globally optimal hyperparameters through the focused nature of the sweet spot. This serves as a critical link between global exploration and local fine-tuning.

[0178] It is worth mentioning that the Starfish optimization algorithm adopts the following position update strategy:

[0179] ;

[0180]

[0181] Where, is the position of the i-th individual at the t-th iteration; is the current optimal individual; is an individual randomly selected from the population to simulate the interference factor; is a random number; is a function that controls the balance between exploration and exploitation, where a and b are constants; represents an exponential decay term with a natural constant as the base.

[0182] S53. The particle swarm optimization algorithm and the gray wolf optimization algorithm of the lower model are used to optimize the optimal domain of the hyperparameters in turn to determine the optimal solution.

[0183] In this embodiment of the present invention, the particle swarm optimization (PSO) algorithm is a heuristic algorithm that simulates the foraging behavior of bird flocks. It randomly initializes a swarm of "particles" (each particle represents a set of hyperparameters) within a hyperparameter optimal domain and iteratively updates the particle positions (hyperparameter values). Each particle adjusts its movement direction based on its own historical best position (pbest) and the global best position (gbest). This algorithm has strong global development capabilities and can quickly converge to a potential optimal region within the optimal domain.

[0184] The Gray Wolf Optimizer (GWO) algorithm simulates the social hierarchy and hunting strategies of wolves. It divides the population into four levels: α, β, δ, and ω. It uses collaboration between these levels to guide search. Wolf α (the current optimal solution) guides the group toward it, while wolves β and δ provide auxiliary search directions, and ω performs random exploration. GWO has stronger local exploration capabilities, allowing it to further refine the search within the area identified by PSO, avoiding being trapped in local optima.

[0185] Particle Swarm Optimization (PSO) rough tuning:

[0186] For each hyperparameter optimal range provided by the upper-layer model, PSO iteratively updates particle positions, rapidly narrowing the search range and focusing on high-potential areas within the optimal range. For example, when optimizing the learning rate and hidden layer size of an LSTM, PSO may converge the search range from [0.001, 0.1] × [64, 256] to [0.005, 0.02] × [128, 192].

[0187] Grey Wolf Optimizer (GWO) fine-tuning:

[0188] Using the PSO output as the initial search point, GWO uses a wolf pack collaboration mechanism to fine-tune hyperparameters. For example, within the small range determined by PSO, GWO might further find the optimal learning rate of 0.008 and hidden layer size of 160, while also optimizing other hyperparameters such as the weight coefficient of the attention mechanism.

[0189] The underlying model dynamically adjusts algorithm weights based on optimization progress. For example, when performance gains slow after multiple iterations, it automatically switches algorithms or combines the advantages of both (such as PSO's global information sharing and GWO's hierarchical guidance) to enhance search flexibility.

[0190] The three-level architecture of "upper-layer global exploration → PSO coarse tuning → GWO fine tuning" avoids the limitations of a single algorithm (such as PSO's tendency to fall into local optima and GWO's low global search efficiency) while significantly reducing computational complexity. The upper-layer model reduces the search space by 10%-30%, while PSO further reduces it by 50%-70%. Ultimately, GWO finds the optimal solution within a very small range, improving overall efficiency by 3-5 times.

[0191] This mechanism can more accurately optimize the model's ability to handle abnormal data in extreme weather scenarios for renewable energy power forecasting. For example, by adjusting the weight coefficients of the extreme scenario sub-model, the model's prediction error in extreme winds or strong radiation conditions can be reduced by 20%-30%.

[0192] It is worth mentioning that the speed and position update formula of the particle swarm optimization algorithm is:

[0193] ;

[0194] ;

[0195] Where, and are the velocity and position of the i-th particle at the t-th iteration, w is the inertia weight, and is the acceleration coefficient, and is a random number, is the historical optimal position of the particle, is the global optimal position; and represent the velocity and position of the i-th particle at the t+1th iteration respectively.

[0196] The position update formula of the gray wolf optimization algorithm is:

[0197] ;

[0198] ;

[0199] ;

[0200] ;

[0201] Where, 、 、 are the position vectors of the three types of leader wolves; represents the new position of the wolf pack individual after the t+1th iteration; 、 、 They are the candidate positions calculated based on the three types of leader wolves; and are coefficient vectors; Indicates the current location of the current individual.

[0202] S54. Adjust the initial prediction model according to the target hyperparameters corresponding to the optimal solution to generate a target prediction model.

[0203] In this embodiment of the present invention, a dynamic hybrid metaheuristic optimization approach, combining global exploration at the upper level and refined optimization at the lower level, yields an optimal solution within the hyperparameter search space that maximizes model performance (e.g., prediction accuracy, stability, etc.). This solution corresponds to a precise set of target hyperparameters (e.g., the optimal number of LSTM network layers, appropriate weights for the attention mechanism, and adaptive parameters for the fully connected layers). These target hyperparameters are based on the performance of the initial prediction model on real-world data (including both common and extreme scenarios), resulting in an algorithmic iteration to identify the optimal configuration. Based on this set of target hyperparameters, the internal structure and parameters of the initial prediction model are modified. For example, if the number of LSTM layers in the optimal solution needs to be adjusted from 2 to 3, the model's recurrent neural network structure is rebuilt; if the attention mechanism weight is adjusted from 0.6 to 0.8, the weight parameters of the corresponding modules are updated. These adjustments free the initial prediction model from the limitations of the original hyperparameter settings and allow it to adapt to data patterns and prediction requirements. Compared with the initial model, the target prediction model generated after adjustment can more accurately capture the complex relationship between time series characteristics, spatial linkage characteristics and extreme weather characteristics when processing new energy power prediction tasks, improve the prediction ability for different scenarios (conventional and extreme), and provide a model tool with better performance and stronger generalization for subsequent reliable new energy power prediction. It is a key transformation step from model optimization exploration to the implementation of practical prediction models.

[0204] It is worth mentioning that before obtaining the target prediction model, the initial prediction model is adjusted to obtain the updated prediction model, and the updated prediction model is further trained and verified. Specifically, the three optimized updated prediction models are trained and predicted respectively, and the updated prediction results are fused using the weighted average method to obtain the final new energy power prediction value. The error between the new energy power prediction value and the actual new energy power value is compared. If it is within the error range, the target prediction model can be determined.

[0205] Specifically, the ensemble learning method can effectively reduce the prediction variance, reduce the risk of overfitting, improve the stability and robustness of the prediction, and further improve the prediction accuracy through the ensemble effect by fusing the prediction results of multiple optimization models and utilizing the complementarity of the models optimized by different algorithms.

[0206] ;

[0207] Where, 、 、 These are the prediction results of the updated prediction model after optimization by particle swarm optimization, grey wolf optimization, and starfish optimization algorithms, respectively. 、 、 are the weights corresponding to the particle swarm optimization, gray wolf optimization and starfish optimization algorithms, and satisfy , the weights are calculated in reverse based on the root mean square error (RMSE) performance of each model. The RMSE indicator is used to evaluate the prediction effect of the model, and the formula is as follows:

[0208] ;

[0209] Where, is the actual power value, is the predicted power value, and N is the number of samples.

[0210] Step 208: Use the target prediction model to predict the new energy power of the new energy power station and generate a target prediction result.

[0211] In the embodiment of the present invention, the specific implementation process of step 208 is similar to that of step 105 and will not be repeated here.

[0212] See also Figure 3 , Figure 3 This is a structural block diagram of a new energy power prediction system provided in Example 3 of the present invention.

[0213] The present invention provides a new energy power prediction system, comprising:

[0214] An acquisition module 301 is used to acquire historical power data of new energy power plants and associated meteorological data and perform preprocessing to generate multi-source data;

[0215] A calculation module 302 is configured to calculate absolute values ​​of correlations between multi-source data and new energy power, and determine a plurality of target multi-source data based on the respective absolute values;

[0216] The prediction module 303 is used to input each target multi-source data into a preset initial prediction model, perform new energy power prediction on each target multi-source data through the initial prediction model, and generate an initial prediction result;

[0217] An optimization module 304 is configured to optimize the hyperparameters of the initial prediction model in sequence using the upper model and the lower model optimized by the dynamic hybrid metaheuristic method based on the initial prediction results to generate a target prediction model;

[0218] The real-time prediction module 305 is used to predict the new energy power of the new energy power station through the target prediction model and generate a target prediction result.

[0219] Furthermore, the acquisition module 301 includes:

[0220] An acquisition submodule is used to acquire historical power data of new energy power stations and associated meteorological data; wherein the new energy power stations include wind power generation and photovoltaic power generation, and the meteorological data includes multi-high-speed wind speed, wind direction, and temperature associated with the historical power data of wind power generation, and solar radiation intensity and ambient temperature associated with the historical power data of photovoltaic power generation;

[0221] The processing submodule is used to perform outlier processing, missing value processing and standardization on historical power data and related meteorological data in sequence to generate multi-source data.

[0222] Furthermore, the system also includes:

[0223] The first judgment submodule is used to judge whether the wind speed of each meteorological data to be measured in the multi-source data is greater than a preset wind speed threshold;

[0224] A first marking submodule is configured to determine that the meteorological data to be measured is extreme weather and mark the meteorological data to be measured as a first extreme weather sample if the weather data to be measured is extreme weather;

[0225] The second judgment submodule is used to judge whether the photovoltaic solar radiation of each meteorological data to be measured in the multi-source data is less than a preset solar radiation threshold if no;

[0226] The second marking submodule is used to determine that the meteorological data to be measured is extreme weather if yes, and mark the meteorological data to be measured as a second extreme weather sample.

[0227] Furthermore, the calculation module 302 includes:

[0228] A calculation submodule, used to calculate the correlation between multi-source data and new energy power based on the Pearson correlation coefficient algorithm;

[0229] An absolute value submodule, for determining the absolute value of the correlation based on the correlation;

[0230] The third judgment submodule is used to judge whether each absolute value is greater than a preset correlation threshold;

[0231] The target multi-source data submodule is configured to determine, if yes, the multi-source data corresponding to the absolute value as the target multi-source data.

[0232] Furthermore, the initial prediction model includes a convolutional layer, a long short-term memory layer, a temporal attention layer, and a fully connected layer; the prediction module 303 includes:

[0233] The input submodule is used to input the target multi-source data into the convolution layer of the preset initial prediction model;

[0234] The convolution submodule is used to convolve the features of the target multi-source data through the convolution layer, generate spatial features and input them into the long short-term memory layer and the temporal attention layer;

[0235] The first extraction submodule is used to extract the effective components of the time dimension and feature dimension of the spatial feature through the long short-term memory layer, generate linkage features and input them into the fully connected layer;

[0236] The second extraction submodule is used to extract the temporal features of the spatial features through the temporal attention layer and input them into the fully connected layer;

[0237] The first connection submodule is used to connect the linkage features and the time series features through the fully connected layer to generate an initial prediction result.

[0238] Furthermore, the initial prediction model also includes an extreme scenario sub-model; the system also includes:

[0239] A screening submodule is used to screen the first extreme weather sample or the second extreme weather sample corresponding to the extreme weather when there is extreme weather in the target multi-source data;

[0240] A third extraction submodule is used to extract extreme weather features of the first extreme weather sample or the second extreme weather sample through the extreme scene submodel and input the features into the fully connected layer;

[0241] The second connection submodule is used to connect the linkage features, time series features and extreme weather features through the fully connected layer to generate the initial prediction results.

[0242] Furthermore, the optimization module 304 includes:

[0243] A search space submodule is used to input the hyperparameter search space of the initial prediction model into the upper model of the dynamic hybrid metaheuristic optimization based on the initial prediction results;

[0244] The exploration submodule is used to perform global exploration of the hyperparameter search space through the starfish optimization algorithm of the upper model, generate multiple hyperparameter optimal domains and input them into the lower model optimized by dynamic hybrid metaheuristics;

[0245] The optimization submodule is used to optimize the hyperparameter optimal range in turn through the particle swarm optimization algorithm and gray wolf optimization algorithm of the lower model to determine the optimal solution;

[0246] The adjustment submodule is used to adjust the initial prediction model according to the target hyperparameters corresponding to the optimal solution to generate the target prediction model.

[0247] A fourth embodiment of the present invention provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed, the new energy power prediction method according to any embodiment of the present invention is implemented.

[0248] Embodiment 5 of the present invention provides a computer program product, which includes a computer program stored on a non-transitory computer-readable storage medium, and the computer program includes program instructions. When the program instructions are executed by a computer, the computer executes a new energy power prediction method as described in any embodiment of the present invention.

[0249] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0250] In the several embodiments provided in this application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are merely schematic. For example, the division of the units is merely a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interfaces, devices or units, which can be electrical, mechanical or other forms.

[0251] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0252] In addition, the functional units in the various embodiments of the present invention may be integrated into a single processing unit, each unit may exist physically separately, or two or more units may be integrated into a single unit. The aforementioned integrated units may be implemented in the form of hardware or software functional units.

[0253] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes various media that can store program code, such as a USB flash drive, a mobile hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0254] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that the technical solutions described in the above embodiments can still be modified, or some of the technical features thereof can be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A new energy power prediction method, characterized in that: include: Obtain historical power data of new energy power plants and associated meteorological data and pre-process them to generate multi-source data; calculating absolute values ​​of correlations between the multi-source data and the new energy power, and determining a plurality of target multi-source data based on the respective absolute values; Inputting each of the target multi-source data into a preset initial prediction model, performing new energy power prediction on each of the target multi-source data using the initial prediction model, and generating an initial prediction result; Based on the initial prediction results, the upper layer model and the lower layer model optimized by dynamic hybrid metaheuristic optimization are used to optimize the hyperparameters of the initial prediction model in turn to generate a target prediction model; The target prediction model is used to predict the new energy power station's new energy power to generate a target prediction result.

2. The new energy power prediction method according to claim 1, characterized in that: The method of obtaining historical power data of new energy power plants and associated meteorological data and preprocessing them to generate multi-source data includes: Obtaining historical power data and associated meteorological data of a new energy power station; wherein the new energy power station includes wind power generation and photovoltaic power generation, and the meteorological data includes multi-high-speed wind speed, wind direction, and temperature associated with the historical power data of wind power generation, and solar radiation intensity and ambient temperature associated with the historical power data of photovoltaic power generation; The historical power data and the associated meteorological data are sequentially processed for outlier value, missing value and standardization to generate multi-source data.

3. The new energy power prediction method according to claim 2, characterized in that: Also includes: Determining whether the wind speed of each meteorological data to be measured in the multi-source data is greater than a preset wind speed threshold; If yes, determining that the meteorological data to be measured is extreme weather, and marking the meteorological data to be measured as a first extreme weather sample; If not, determining whether the photovoltaic solar radiation of each meteorological data to be measured in the multi-source data is less than a preset solar radiation threshold; If so, it is determined that the meteorological data to be measured is extreme weather, and the meteorological data to be measured is marked as a second extreme weather sample.

4. The new energy power prediction method according to claim 1, characterized in that: The calculating the absolute value of the correlation between the multi-source data and the new energy power, and determining a plurality of target multi-source data based on each of the absolute values, includes: Calculating the correlation between the multi-source data and the new energy power based on the Pearson correlation coefficient algorithm; determining an absolute value of the correlation based on the correlation; Determining whether each of the absolute values ​​is greater than a preset correlation threshold; If so, the multi-source data corresponding to the absolute value is determined as the target multi-source data.

5. The new energy power prediction method according to claim 3, characterized in that: The initial prediction model includes a convolutional layer, a long short-term memory layer, a temporal attention layer, and a fully connected layer; the target multi-source data are input into a preset initial prediction model, and the new energy power of each target multi-source data is predicted by the initial prediction model to generate an initial prediction result, including: Inputting each of the target multi-source data into the convolutional layer of a preset initial prediction model; Convolving the features of the target multi-source data through the convolution layer to generate spatial features and input them into the long short-term memory layer and the temporal attention layer; Extracting effective components of the time dimension and feature dimension of the spatial feature respectively through the long short-term memory layer, generating linkage features and inputting them into the fully connected layer; Extracting the temporal features of the spatial features through the temporal attention layer and inputting them into the fully connected layer; The linkage feature and the time series feature are connected through the fully connected layer to generate an initial prediction result.

6. The new energy power prediction method according to claim 5, characterized in that: The initial prediction model also includes an extreme scenario sub-model; the method further includes: When extreme weather exists in the target multi-source data, screening a first extreme weather sample or a second extreme weather sample corresponding to the extreme weather; Extracting extreme weather features of the first extreme weather sample or the second extreme weather sample through the extreme scene sub-model and inputting the features into the fully connected layer; The linkage feature, the time series feature, and the extreme weather feature are connected through the fully connected layer to generate an initial prediction result.

7. The new energy power prediction method according to claim 1, characterized in that: Based on the initial prediction result, the upper layer model and the lower layer model optimized by the dynamic hybrid metaheuristic method are sequentially optimized to optimize the hyperparameters of the initial prediction model to generate a target prediction model, including: Based on the initial prediction results, the hyperparameter search space of the initial prediction model is input into the upper model of the dynamic hybrid metaheuristic optimization; Performing a global exploration of the hyperparameter search space using the starfish optimization algorithm of the upper model to generate multiple hyperparameter optimal domains and inputting them into the lower model optimized by the dynamic hybrid metaheuristic method; Optimizing the hyperparameter optimal range in turn by the particle swarm optimization algorithm and the gray wolf optimization algorithm of the lower model to determine the optimal solution; The initial prediction model is adjusted according to the target hyperparameters corresponding to the optimal solution to generate a target prediction model.

8. A new energy power prediction system, characterized in that: include: The acquisition module is used to obtain historical power data of new energy power plants and related meteorological data and perform preprocessing to generate multi-source data; a calculation module, configured to calculate absolute values ​​of correlations between the multi-source data and the new energy power, and determine a plurality of target multi-source data based on the respective absolute values; A prediction module, configured to input each of the target multi-source data into a preset initial prediction model, perform new energy power prediction on each of the target multi-source data using the initial prediction model, and generate an initial prediction result; An optimization module is used to optimize the hyperparameters of the initial prediction model in sequence using an upper model and a lower model optimized by a dynamic hybrid metaheuristic method based on the initial prediction result to generate a target prediction model; The real-time prediction module is used to predict the new energy power of the new energy power station through the target prediction model and generate a target prediction result.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed, the new energy power prediction method according to any one of claims 1 to 7 is implemented.

10. A computer program product, characterized in that The computer program product includes a computer program stored on a non-transitory computer-readable storage medium, wherein the computer program includes program instructions, wherein when the program instructions are executed by a computer, the computer is caused to execute the new energy power prediction method according to any one of claims 1 to 7.

Images